When do LLM agents develop new languages that we can’t understand?
Lots of recent news about this, based mostly on anecdata from a single run. We study language emergence more rigorously, finding key factors like LLM strength, access to scratchpad messages, and pressure for efficiency.
Studying the languages themselves, we find they are morphologically productive, compositional, and can be transmitted to new agents, including agents backed by weaker models, even ones not able to develop language on their own.
To study language emergence systematically, we developed a new platform, GlossoGen, which lets us design controlled, sandboxed multi-agent scenarios with different initial conditions and dynamics. We instantiate one such scenario and use it to study open and closed-weight models across many runs.
Key takeaways:
1️⃣ Sufficiently strong models, under pressure to communicate efficiently and with access to a postmortem scratchpad, develop new languages.
2️⃣ Languages are compositional and morphologically productive.
3️⃣ Languages can be transmitted to new learners who observe them being used without seeing their construction.
4️⃣ Even models that are not strong enough to construct languages can learn to use them. Agents take an active role in learning languages, with new agents repairing failed conversations via targeted queries.
More details in our paper below, including implications for safety/monitorability, cumulative cultural evolution, and linguistics.
🧵👇
3
38
135
10,995
Studying agent language emergence requires a new environment where we can run repeatable, controlled experiments, rewind simulations, and play out counterfactuals.
We created GlossoGen for this purpose. GlossoGen lets you design scenarios to study language emergence between agents, supporting a variety of agent roles, different environment events and pressures, and dedicated communication channels, all in a safe, controlled sandbox where agents cannot execute code.
1
1
4
425
The first scenario we designed mimics an emergency response scenario, in which an EMT needs to communicate with a doctor to save a patient. In our case, the patient is Veyru, a fictional alien (this prevents LLMs from relying on parametric medical knowledge).
The scenario has information asymmetry, with the “doctor” agent having access to the remedy to Veyru’s symptoms, while the “EMT” agent observes and communicates those symptoms.
Agents are under pressure to communicate efficiently, with a fixed budget of characters they can transmit. They also have access to a postmortem stage, in which they can communicate without budget restrictions about the previous round’s performance.
Sep 3, 2026 · 3:20 PM UTC
1
1
4
365
Finding 1: Agents develop new languages that do not look like English
Under the right conditions, agents consistently develop diverse new languages that are uninterpretable to a human, as reflected in increased perplexity under an English LM (GPT-2). This has important implications for monitorability: if we can’t understand the language agents use then we can’t monitor their performance, and can only inspect outcomes.
Critically, these languages differ between runs, even of the same model.
1
1
4
288
Finding 2: Three factors govern emergence: model strength, access to a postmortem stage, and pressure.
Frontier models (GPT 5.4, Sonnet 4.6) develop new languages when subject to budget pressure, but not when given an unlimited budget. Open-source models (Qwen3 32B, Llama3-70B) fail to develop new languages. Agents develop new languages when they have access to postmortem stages, but fail to do so without.
1
1
4
301
Finding 3: Languages are morphologically productive.
We looked at the morphology of these new languages and found that they are compositional and productive: agents can correctly produce and parse new forms that are licensed under the slots of the language’s morphosyntax, but have never been observed in the agent’s history.
The morphosyntax is productive and sometimes differs from English – for example, some languages use capitalization contrastively (likely, a result of our per-character cost), and word order often differs from English.
1
1
4
213
Finding 4: Cumulative cultural evolution: agents transmit languages to newcomer agents.
We swap in new agents who have not seen the postmortem stages and must learn from usage alone. In many cases, these agents learn the new language, and some languages are more easily learned than others. This variable transmissibility is a critical ingredient for cumulative cultural evolution, where agents may iteratively evolve a technology (in this case, language) over multiple generations.
Interestingly, while open-source models do not develop their own languages, they are sometimes able to learn to use existing languages that other agents have constructed, indicating that language innovation may only require a few “special” agents, but could affect a whole population.
Agents also take an active role in this transmission process, engaging in conversational repair to recover the meaning of expressions they don’t know. They query more on non-compositional terms, instead inferring the meaning of compositional terms from their parts.
1
1
5
212
Lots more to unpack in the paper! Some exciting and important discussion points we raise:
1️⃣ Safety: Our scenario is purely cooperative, and we still see agents developing new languages that would be hard to monitor or audit. By documenting the variables that lead to this emergence, we hope our work can further research into how to maintain interpretable multi-agent systems at scale.
2️⃣ Cumulative cultural evolution (CCE): Our results suggest agents already have the ability to perform CCE. This has important implications for how we think of AI, in terms of the capabilities of the individual vs. aggregate capability of the system. Language is an important part of this: it’s both a product of CCE and a technology that enables further CCE.
3️⃣ Linguistics: I'm bullish about this as a tool to study language, and equally about linguistics as a tool to analyze multi-agent language. A lot of work has looked at language in LLMs in a “synchronic” way (one snapshot in time). This gives us a new “diachronic” lens through which to study the linguistic generalizations of LLMs: how a model’s language changes over time can tell us something about how the model represents language, and about the cognitive capabilities of the model. Our morphological analyses also highlight the role of linguistics as the right toolkit for decoding and analyzing agent language, a toolkit we may find ourselves increasingly in need of.
1
1
5
250
This was a really fun collaborative project, spurred on by @schmidtsciences, combining talents across many institutions, with Newton Sander, Carlos Bonetti, @SashaBoguraev, James Bowler, Hale Sirin, and @SimonKirby
@aestudio @UTCompSci @UTAustin
Paper: arxiv.org/abs/2609.01491
4
8
383